Practical guide

AI account reconciliation verification checklist

Use this checklist after preparing an AI-assisted account reconciliation and before an authorized reviewer releases it. The checklist tests whether the balance can be reperformed from preserved sources, whether every difference has an evidence-backed disposition, and whether the model's contribution stayed inside its permitted role. A checked box is not proof by itself. The evidence reference,.

By Two PruneUpdated 10/15/2026

Set the reconciliation boundary before reviewing conclusions

Record the entity, account, currency, period end, ledger extract, independent comparison source, cut-off rule, preparer, and reviewer. If any field is missing, stop and return the schedule for repair. An apparently balanced schedule can still be unreliable when it compares different periods, entities, or source populations.

The 2025 GAO Green Book treats internal control as a management process and emphasizes risk-responsive control activities, reliable information, documented performance, and ongoing evaluation. It became effective for U.S. federal fiscal year 2026. This checklist uses those control ideas as educational design guidance; it does not claim that applying them makes a private reconciliation compliant with the Green Book.

[ ] The ledger extract has an immutable file name or report ID and extraction timestamp.
[ ] The independent balance comes directly from its authoritative system or custodian.
[ ] The period, currency, entity, and account agree across both sources.
[ ] The cut-off treatment is written, including late postings and in-transit items.
[ ] The preparer and reviewer are different people where the control design requires separation.

Sources: [1]

Trace every input to a source that a reviewer can reopen

Do not accept pasted totals, screenshots without context, or an AI summary as the source record. For each balance and reconciling item, preserve a reference that another authorized person can reopen. Record where the evidence came from, who obtained it, and whether it was transformed before review.

The current PCAOB auditing standards compilation distinguishes evidence about control design from evidence that a control operated, and requires sufficient appropriate evidence for audit conclusions. A management reconciliation is not an audit, but the distinction is useful: a beautifully designed checklist does not show that any item was actually checked.

[ ] Opening and closing balances trace to named source rows or statement pages.
[ ] Each reconciling item has a stable item ID and evidence reference.
[ ] AI-visible data is identified, including masking, filtering, and excluded fields.
[ ] Manual adjustments are distinguished from system-posted transactions.
[ ] Evidence access was tested by the reviewer, not assumed from a hyperlink label.

Sources: [2]

Reperform the arithmetic without relying on the narrative

Start from the preserved source balances and recompute the tie-out. Keep formulas visible. Never accept a model-generated explanation because it sounds coherent or because a final difference displays zero. The arithmetic must work row by row, and sign conventions must remain consistent.

When AI was used to write formulas or classify transactions, record the prompt or instruction, the output version, and the human check. AI may propose a category or calculation. It must not silently post an entry, replace the source record, or approve the result.

[ ] The reviewer recalculated the source-to-ledger bridge independently.
[ ] Debits, credits, positive values, and negative values use one documented convention.
[ ] No unrelated items were netted to conceal offsetting errors.
[ ] Duplicate rows, hidden rows, filters, and hard-coded formula results were checked.
[ ] Rounding differences have an approved threshold and explicit treatment.
[ ] The final adjusted balances agree, or the schedule is marked hold.

Test each exception for evidence, age, ownership, and authority

Every open item needs a distinct disposition. Use cleared only when evidence resolves the item and no action remains. Use repair when the cause is understood but a posting, confirmation, or document is outstanding. Use hold when the cause is unknown, evidence is insufficient, or the required authority has not approved the treatment.

The reviewer should sample both high-risk and ordinary items. Only reviewing the largest difference can miss systematic issues such as repeated fees, duplicate mappings, or a broken source feed.

[ ] Each exception has amount, date, category, owner, due date, and status.
[ ] The proposed treatment cites evidence rather than repeating the transaction description.
[ ] Aged or material items follow the organization's escalation rule.
[ ] Rejected AI suggestions remain visible in the review trail.
[ ] Proposed journal entries identify the authorized poster and approver.
[ ] Unresolved contradictions are not converted into vague monitoring notes.

Worked checklist: a reconciliation that must remain on hold

Assume a synthetic reconciliation compares a ledger closing balance of SGD 248,400 with a bank statement closing balance of SGD 252,950. The schedule lists a SGD 6,000 deposit in transit, a SGD 1,200 bank fee, a SGD 3,500 duplicate ledger payment, and a SGD 150 unexplained statement credit. An AI assistant labels the deposit and duplicate payment as timing items.

The boundary and source checks pass: the files cover the same entity, account, currency, and 30 September cut-off. The reviewer reopens the statement and ledger extract, then rejects the duplicate-payment label after finding two ledger postings tied to one approved instruction. The arithmetic check produces an unexplained remaining gap of SGD 3,250 rather than the zero claimed in the draft narrative. The checklist records the rejected AI suggestion, links the duplicate entries, assigns the fresh-extract request to the controller, and marks the entire reconciliation hold.

This is the expected behavior. The checklist prevents a fluent explanation and a plausible category from overriding source evidence. The worked artifact consists of the checked list, source index, visible bridge calculation, exception register, reviewer decision, and retained AI contribution log.

Record the release decision and residual uncertainty

The reviewer should choose release, repair, or hold and explain the decision in one bounded paragraph. Record the reviewed version, date, reviewer, unresolved items, next action, and approval authority. If the schedule is released with a tolerated difference, state the threshold, policy owner, and why the remaining uncertainty is acceptable for this specific purpose.

Do not describe completion of this checklist as proof that financial statements are correct, that a control is effective across periods, or that legal, accounting, or audit requirements have been met. Those conclusions require the applicable framework, qualified judgment, and broader evidence.

Two Prune can use a synthetic reconciliation and this checklist inside a realistic assessment to observe how a participant frames the task, navigates evidence, uses AI, verifies calculations, escalates uncertainty, and communicates a bounded conclusion. Two Prune does not perform the reconciliation, approve balances, certify controls, rank people, or make employment decisions.

[ ] The disposition matches the status of every material or aged exception.
[ ] Review evidence includes identity, timestamp, version, and scope.
[ ] Residual uncertainty is stated rather than implied away.
[ ] Follow-up ownership and due dates are visible.
[ ] The record is retained according to the applicable policy.

Sources

1.2025 GAO Green Book · U.S. Government Accountability Office
2.PCAOB auditing standards compilation · Public Company Accounting Oversight Board